B.-J. Falkowski (Germany)
Machine Learning, large margin pocket algorithm, kernel methods, perceptron learning, ranking problems.
The Large Margin Kernel Pocket Algorithm as described by Xu et al. represents an interesting alternative to support vector machines. Unfortunately though its theoretical basis appears slightly shaky. Here under a weak assumption a shorter proof of a known result is given. This leads to the discovery of a gap in the proof of the mentioned algorithm. Also for the case of a simple feature map non-negligible CPU times are required as tests with a larger number of samples show. Hence and because of its practical significance an explicit dual of a slightly modified version, the Large Margin Pocket Algorithm, is given which provides superior run times for simple feature maps. In spite of there being no guarantee for optimal solutions it should be of practical relevance in view of preliminary test results.
Important Links:
Go Back
